TB-6275 · REV E · Technical newsheet
Sensors & InstrumentationDevice profile
Torso chrominance cuts breathing-rate error to 1.267 breaths/min
A Sensors study finds an untrained upper-torso chrominance channel estimates breathing rate at 1.267 breaths/min MAE, beating facial colour channels and multi-stream networks.
By Olivia Hart4 min read753 words
Features
- Untrained upper-torso chrominance channel reached MAE 1.267 breaths per minute and correlation +0.795
- Facial colour channels produced errors of 3.693 to 3.912 breaths/min, worse than the ~2.644 baseline
- Five stationary background patches gave 4.359 breaths/min; neck region 2.580; sub-pixel row displacement 2.917
- Removing in-band chest motion from chrominance lowered error to 1.156 breaths/min
- Evaluation used COHFACE (160 clips, 39 subjects) and VitalVideos, with 0.10–0.50 Hz Butterworth bandpass and five-fold subject-disjoint validation
- Published as Sensors 26(22), DOI 10.3390/s26227107, 2026

The strongest respiratory-rate estimate in a camera-only comparison came from a single untrained colour channel over the upper torso: mean absolute error (MAE) of 1.267 breaths per minute, correlation coefficient +0.795. No neural network, no training set, no chest strap.
Buyung, R. A. and colleagues published the comparison as "Colour-Channel Selection for Video-Based Respiratory Rate Estimation: A Diagnostic Study with a Single-Shot Held-Out Evaluation" in Sensors, volume 26, issue 22 (2026), DOI 10.3390/s26227107. Their conclusion carries a purchasing implication: anatomical region and colour-channel selection matter more than model complexity.
Why do facial video signals lose the breathing rate?
Remote photoplethysmography (rPPG) extracts vital signs from video by tracking subtle skin-colour shifts caused by blood-volume pulses. Scientists built rPPG for heart rate and pulse variability, then found respiration modulates the same optical channel through changes in skin colour, vascular volume, and rhythmic body movement.
Facial regions dominate the literature because face-detection software crops faces reliably. Chest and abdominal excursion, the mechanical indicator of breathing, has drawn less attention. Buyung's group asked which region, and which colour projection, actually resolves the respiratory waveform.
How was the comparison built?
Two datasets anchored the evaluation. COHFACE supplied 160 video clips from 39 subjects with reference measurements from respiration-belt sensors. VitalVideos contributed high-frame-rate biomedical recordings.
Three colour projections were tested: chrominance-based (CHROM), plane-orthogonal-to-skin (POS), and the green-red-blue difference method (2G-R-B). Each signal came from the face and from an upper-torso bounding box defined relative to the face, while grayscale spatial patches tracked movement.
The pipeline ran on Python 3.10 with NumPy, SciPy, OpenCV, and MediaPipe for facial landmark detection. Clips were cut into fixed windows, filtered through a Butterworth bandpass covering 0.10–0.50 Hz, and resampled to 4 Hz. A windowed periodogram located the dominant spectral peak. Five-fold, subject-disjoint validation kept the same participants out of both training and testing partitions. Rather than large deep networks, the team compared compact dual-stream models driven by deterministic, training-free colour signals.
What did the face lose to?
Facial colour channels produced errors from 3.693 to 3.912 breaths per minute, all worse than the baseline of roughly 2.644 breaths per minute. Inside the torso region, only the chrominance projection reliably separated breathing from background noise; the other colour representations failed.
Was the torso signal just lighting or chest motion?
Five stationary background patches returned an error of 4.359 breaths per minute, so the fluctuations did not explain the signal. Shifting the region upward to the neck raised the error to 2.580 breaths per minute, which argues for deliberate torso placement rather than an arbitrary crop.
A separate mechanism test asked whether chrominance merely mirrored physical movement. Full-resolution sub-pixel row displacement scored an error of 2.917 breaths per minute. Stripping the in-band mechanical motion component out of the chrominance signal dropped the error to 1.156 breaths per minute, below the original 1.267. Respiratory information in that channel therefore resists explanation by chest motion alone.
Where could a lightweight system sit?
Buyung's team points to camera-based monitoring in hospitals, telemedicine, and smart home environments. Because the primary respiratory signal comes from an untrained chrominance stream over the upper torso, a monitor could run without intensive deep learning, on edge computing devices and low-cost optical sensor modules. The authors name neonatal incubators and contactless bedside monitors as candidates.
The torso also offers a privacy route: developers can blur or mask facial features while retaining the torso region for physiological measurement. That shifts the anonymity trade-off away from the face, where identity recognition and clinical detail overlap.
What still has to be proven before adoption?
The study measured datasets, not wards. The authors call for work in active clinical environments, with particular attention to skin tone variation, fabric texture, and movement, and note that improved optical motion estimation may address part of that problem. They also argue that consistent benchmark protocols will determine whether camera-based physiological monitoring earns clinical trust.
The open question for buyers is narrow and answerable: does a 1.267 breaths per minute mean absolute error survive beside a breathing infant in an incubator, or a moving patient in a bed, under the lighting and clothing of an actual ward? A torso chrominance stream now looks like the cheapest candidate to test first.
via azosensors.com (Source)
Filed under
- respiratory-rate
- remote-photoplethysmography
- chrominance
- camera-based-monitoring
- contactless-monitoring
More from Olivia Hart
Show full bio
Market editor covering technology and product news in Instruments, test and measurement at Testbench Report.
97 articles
Application notes
- Industrial IoT Cameras Demand Sensor Rework for the AI Era
- Human Inspection Accuracy Falls 20–30% After One Hour
- Machine Vision System Targets Online Fabric Color Difference Detection
- Machine Vision as the Optical Floor of AI-Driven Automation
- Quality Magazine: Machine Vision Moves to On-Camera Inference